All pull requests

Pull requests

• Loosen the late-iteration gradient tolerance in the checkpointing test

  #8648·delock opened Sep 24, 2026Collaborator·

  ·

• Honor fp16 hysteresis in the ZeRO stage 0 fp16 optimizers

  #8647·vineethsaivs opened Sep 23, 2026Contributor·

  ·

• Fix NameError in the MoE socket and machine checkpoint writer

  #8646·vineethsaivs opened Sep 23, 2026Contributor·

  ·

• Add an opt-in fused Triton RMSNorm for Llama-style Hugging Face models

  #8645·yh0903 opened Sep 23, 2026Contributor·

  ·

• Add opt-in fused AutoEP DeepEP row weighting

  #8644·yh0903 opened Sep 23, 2026Contributor·

  ·

• Link every param group's optimizer state in ZeRO-1/2

  #8643·alanhuangyoo opened Sep 23, 2026Contributor·

  ·

• Fix scalar parameter byte-view broadcasts

  #8642·adenzhou1350 opened Sep 23, 2026·

  ·
  1

• Fix Muon aspect-ratio scale truncated when muon_update recompiles

  #8641·alanhuangyoo opened Sep 23, 2026Contributor·

  ·
  2

• [BUG] Replace deprecated npu_apply_adam(_w) with torch._fused_adam(w)_ on NPU

  #8640·Ginray opened Sep 23, 2026·

  ·

• Keep invalid group-norm sentinels from becoming a finite 1.0 step

  #8638·sfc-gh-truwase opened Sep 23, 2026Collaborator·

  ·

• Link ZeRO-1/2 fragments at their round-robin offsets

  #8637·alanhuangyoo opened Sep 23, 2026Contributor·

  ·

• Replay AutoEP's cached DeepEP dispatch in the layout it was built with

  #8636·yh0903 opened Sep 23, 2026Contributor·

  ·

• Use expert gradients, not expert weights, in the FP16_Optimizer MoE grad norm

  #8635·yh0903 opened Sep 23, 2026Contributor·

  ·

• [MoE] Add AutoEP routing replay

  #8634·Silas-Zeng opened Sep 23, 2026·

  ·

• Stage Muon's momentum once per partition in ZeRO-1/2

  #8633·alanhuangyoo opened Sep 23, 2026Contributor·

  ·

• Harden ZeRO-2 CPU-offload gradient storage, synchronization, and accumulation

  #8632·jeffra opened Sep 23, 2026Collaborator·

  ·
  2

• Fix AutoSP tests that expect LLM attention to be wrapped with DistributedAttention

  #8631·ankitdhandharia opened Sep 22, 2026·

  ·
  1

• [MoE] Validate grouped AutoEP top-k routing

  #8630·Silas-Zeng opened Sep 22, 2026·

  ·

• Fix ZeRO-3 Muon all-gather for mixed parameter shapes

  #8628·jinyouzhi opened Sep 22, 2026Contributor·

  ·
  3

• Implement FusedLion, FusedLamb, and Quantizer ops for NPU accelerator

  #8627·VenusTZZ opened Sep 22, 2026Contributor·

  ·
  1

• Avoid repeated full gradient clearing in ZeRO-3 subgroup updates

  #8625·aprylewu opened Sep 22, 2026·

  ·
  1

  2

• [checkpoint] Refuse a map whose pieces write the same shard address

  #8624·0z5a opened Sep 22, 2026Contributor·

  ·

• [UC] Plan a scale-1 affine shard-to-shard transfer without rebuilding the parameter

  #8623·0z5a opened Sep 22, 2026Contributor·

  ·
  2

• Restore a shard from the map the layer published

  #8622·Achyuthan-S opened Sep 21, 2026Contributor·

  ·
  3

• Stop top-2 MoE gating from sampling the second expert in eval

  #8620·vineethsaivs opened Sep 21, 2026Contributor·

  ·